A master thesis from Aalborg University
Deep Learning Based Adversarial Examples Detection for Speech Recognition: A Comparative Study of Speech Processing Methods
Term
4. term
Education
Publication year
2019
Submitted on
2019-06-05
Abstract
This project investigates the use of neural networks for detecting adversarial examples for speech recognition with five different feature extraction methods as input. These are STFT, MFCC, IMFCC, GFCC and IGFCC. Relevant theory in the areas of deep learning, adversarial examples and speech processing is examined, and a description is made of the available white box and black box datasets. A CNN model is implemented and evaluated w.r.t performance and robustness of the different feature extraction methods. This includes an investigation into how the performance is affected when only speech or nonspeech is present in the data. The addition of different types and amounts of noise in the data is investigated to determine how it affects the performance of the model. It is concluded that the CNN model is able to detect adversarial examples for speech recognition and the IMFCC and IGFCC feature extraction methods in general have highest accuracies. Furthermore the model is generally more robust to noise when its training set contains a wider range of noise types.
Documents
Colophon: This page is part of the AAU Student Projects portal, which is run by Aalborg University. Here, you can find and download publicly available bachelor's theses and master's projects from across the university dating from 2008 onwards. Student projects from before 2008 are available in printed form at Aalborg University Library.
If you have any questions about AAU Student Projects or the research registration, dissemination and analysis at Aalborg University, please feel free to contact the VBN team. You can also find more information in the AAU Student Projects FAQs.